Time series data storage in Redis, Mongo, SQL and Cassandra
Project description
- Version:
- 0.8.0
- Download:
- Source:
- Keywords:
python, redis, mongo, sql, mysql, sqlite, postgresql, time, timeseries, rrd, gevent, statistics
Overview
Kairos provides time series storage using Redis, Mongo, SQL or Cassandra SQL backends. Kairos is intended to replace RRD and Whisper in situations where the scale and flexibility of other data stores is required. It works with gevent out of the box. Kairos is the library on which torus is built.
Recommended for python 2.7 and later, it can work with previous versions if you install OrderedDict.
Usage
Kairos supports all storage engines using the same API. The constructor will return a Timeseries instance tailored for the type of data and the storage engine, and the API for updating and querying the timeseries is consistent for all combinations of data type and storage engine.
Constructor
The first argument is a handle to a supported storage engine, and the rest of the keyword arguments configure the timeseries. The keyword arguments supported by all storage engines are:
type
Optional, defaults to "series". Can be one of:
series - each interval will append values to a list
histogram - each interval will track count of unique values
count - each interval will maintain a single counter
gauge - each interval will store the most recent data point
set - each interval will store a set of unique values
read_func
Optional, is a function applied to all values read back from the
database. Without it, values will be strings for Redis, whatever
`write_func` defined for Mongo. Must accept a string value for Redis
(empty string for no data) and can return anything.
write_func
Optional, is a function applied to all values when writing. Can be
used for histogram resolution, converting an object into an id, etc.
Must accept whatever can be inserted into a timeseries and return an
object which can be saved according to the rules of the storage engine.
intervals
Required, a dictionary of interval configurations in the form of:
{
# interval name, used in keys and should conform to best
# practices according to the storage engine.
minute: {
# Required. The number of seconds that the interval will cover,
# or a supported Gregorian interval.
step: 60,
# Optional. The maximum number of intervals to maintain. If supplied,
# will use Redis and Mongo expiration to delete old intervals, else
# intervals exist in perpetuity. If the storage engine doesn't support
# expiry, will be used to implement the expire() call.
steps: 240,
# Optional. Defines the resolution of the data, i.e. the number of
# seconds in which data is assumed to have occurred "at the same time".
# So if you're tracking a month-long time series, you may only need
# resolution down to the day, or resolution=86400. Defaults to same
# value as "step". Can also be a Gregorian interval.
resolution: 60,
}
}
In addition to specifying step and resolution in terms of seconds, kairos also supports a simplified format for larger time intervals. For hours (h), days (d), weeks (w), months (m) and years (y), you can use the format 30d to represent 30 days, for example.
As of 0.3.0, kairos also supports the Gregorian calendar for step and resolution. Either or both parameters can use the terms [daily, weekly, monthly, yearly] to describe an interval. You can also mix these terms between step and resolution (e.g. daily in 1h resolutions). The expiration time for Gregorian dates is still defined in terms of seconds and may not match the varying month lengths, leap years, etc. Gregorian dates are translated into strptime- and strftime-compatible keys (as integers) and so may be easier to use in raw form or with any external tools.
Each retrieval function will by default return an ordered dictionary, though condensed results are also available. Run script/example to see standard output; watch -n 4 script/example is a useful tool as well.
Storage Engines
Each of the supported storage engines also supports a set of keyword arguments to configure their behavior.
Redis
An example timeseries stored in Redis:
from kairos import Timeseries
import redis
client = redis.Redis('localhost', 6379)
t = Timeseries(client, type='histogram', read_func=int, intervals={
'minute':{
'step':60, # 60 seconds
'steps':120, # last 2 hours
}
})
t.insert('example', 3.14159)
t.insert('example', 2.71828)
print t.get('example', 'minute')
Additional keyword arguments are:
prefix Optional, Redis only, is a prefix for all keys in this timeseries. If supplied and it doesn't end with ":", it will be automatically appended.
Mongo
An example timeseries stored in Mongo:
from kairos import Timeseries
import pymongo
client = pymongo.MongoClient('localhost')
t = Timeseries(client, type='histogram', read_func=int, intervals={
'minute':{
'step':60, # 60 seconds
'steps':120, # last 2 hours
}
})
t.insert('example', 3.14159)
t.insert('example', 2.71828)
print t.get('example', 'minute')
SQL
An example timeseries stored in a SQLite memory store:
from kairos import Timeseries
from sqlalchemy import create_engine
client = create_engine('sqlite:///:memory:')
t = Timeseries(client, type='histogram', read_func=int, intervals={
'minute':{
'step':60, # 60 seconds
'steps':120, # last 2 hours
}
})
t.insert('example', 3.14159)
t.insert('example', 2.71828)
print t.get('example', 'minute')
Additional keyword arguments are:
string_length Optional, configures the length of strings (VARCHARs). Defaults to 255. All tables have at least 2 string columns, and the size of these columns may impact usability of the SQL storage engine. text_length Optional, configures the length of TEXT and BLOB columns. Defaults to 32Kbytes. Only matters if value_type is a text or blob. table_name Optional, overrides the default table name for a timeseries type. value_type Optional, defines the type of value to be stored in the timeseries. Defaults to float. Can be a string, a Python type or a SQLAlchemy type or instance. 'blob' 'bool' <type 'bool'> 'boolean' 'clob' 'date' <type 'datetime.date'> 'datetime' <type 'datetime.datetime'> 'decimal' <class 'decimal.Decimal'> 'float' <type 'float'> 'int' 'int64' 'integer' <type 'int'> 'long' <type 'long'> 'str' 'string' <type 'str'> 'text' 'time' <type 'datetime.time'> 'unicode' <type 'unicode'>
Cassandra
An example timeseries stored in Cassandra:
from kairos import Timeseries
import cql
client = cql.connect('localhost', 9160, 'keyspace', cql_version='3.0.0')
t = Timeseries(client, type='histogram', read_func=int, intervals={
'minute':{
'step':60, # 60 seconds
'steps':120, # last 2 hours
}
})
t.insert('example', 3.14159)
t.insert('example', 2.71828)
print t.get('example', 'minute')
Additional keyword arguments are:
table_name Optional, overrides the default table name for a timeseries type. value_type Optional, defines the type of value to be stored in the timeseries. Defaults to float. Can be a string or a Python type. <type 'unicode'> string decimal <type 'long'> int double unicode float long <type 'bool'> <type 'float'> boolean int64 str text blob clob integer bool <type 'str'> <type 'int'> inet
kairos requires cql as it supports CQL3 and gevent. This requires that the keyspace be created before the connection, and the keyword argument cql_version='3.0.0' must be used.
A notable downside of this library is that it does not support a list of endpoints to connect to, so is missing key High Availability features.
It is likely that future versions of kairos will require cassandra-driver when it is ready.
Cassandra counters can only store integers, and cannot be used for a running total of floating point numbers.
Inserting Data
There is one method to insert data, Timeseries.insert.
insert
name The name of the statistic
value The value of the statistic (optional for count timeseries)
timestamp (optional) The timestamp of the statistic, defaults to time.time() if not supplied
For series and histogram timeseries types, value can be whatever you’d like, optionally processed through the write_func method before being written to storage. Depending on your needs, value (or the output of write_func) does not have to be a number, and can be used to track such things as unique occurances of a string or references to other objects, such as MongoDB ObjectIds. Note that many of the aggregate functions in histogram expect the data to be real numbers.
For the count type, value is optional and should be a float or integer representing the amount by which to increment or decrement name; it defaults to 1.
For the gauge type, value can be anything and it will be stored as-is.
Data for all timeseries is stored in “buckets”, where any Unix timestamp will resolve to a consistent bucket name according to the step and resolution attributes of a schema. A bucket will contain the following data structures for the corresponding series type.
series list
histogram dictionary (map)
count integer or float
gauge value
Discovering Data
There are two methods to “discover” the data store in a Timeseries.
list
There are no arguments. Returns a list of all of the stat names stored in the Timeseries.
properties
Takes a single argument, the name of the timeseries. Returns a dictionary with the following fields:
{ interval : { 'first' : timestamp, 'last' : timestamp } }
interval will be the named interval, such as “minute”. For each interval, there is a dictionary of properties. first is the timestamp of the first data point in the timeseries, and last is the last data point in the timeseries.
Reading Data
There are three methods to read data, Timeseries.get, Timeseries.series and Timeseries.iterate. get will return data from a single bucket, and series will return data from several buckets. iterate will use the Timeseries.properties method to determine the date range of the data, and return a generator that calls get for every possible interval in the date range.
get
Supports the following parameters. All optional parameters are keyword arguments.
name The name of the statistic, or a list of names whose data will be joined together.
interval The named interval to read from
timestamp (optional) The timestamp to read, defaults to time.time()
condensed (optional) DEPRECATED Use condense instead. Support for this will be removed entirely in a future release.
transform (optional) Optionally process each row of data. Supports [mean, count, min, max, sum], or any callable that accepts datapoints according to the type of series (e.g histograms are dictionaries, counts are integers, etc). Transforms are called after read_func has cast the data type and after resolution data is optionally condensed. If transform is one of (list,tuple,set), will load the data once and run all the transforms on that data set. If transform is a dict of the form { transform_name : transform_func }, will run all of the transform functions on the data set.
fetch (optional) Function to use instead of the built-in implementations for fetching data. See Customized Reads.
process_row (optional) Can be a callable to implement Customized Reads.
condense (optional) If using resolutions, True will collapse the resolution data into a single row. Can be a callable to implement Customized Reads.
join_rows (optional) Can be a callable to implement Customized Reads.
Returns a dictionary of { timestamp : data }, where timestamp is a Unix timestamp and data is a data structure corresponding to the type of series, or whatever transform returns. If not using resolutions or condensed=True, the length of the dictionary is 1, else it will be the number of resolution buckets within the interval that contained data. If transform is a list, data will be a dictionary of { transform_func : transformed_data }. If transform is a dict, data will be a dictionary of { transform_name : transformed_data }.
series
Almost identical to get, supports the following parameters. All optional parameters are keyword arguments.
name The name of the statistic, or a list of names whose data will be joined together.
interval The named interval to read from
start (optional) The timestamp which should be in the first interval of the returned data.
end (optional) The timestamp which should be in the last interval of the returned data.
steps (optional) The number of steps in the interval to read, defaults to either steps in the configuration or 1. Ignored if both start and end are defined. If either start or end are defined, steps is inclusive of whatever interval that timestamp falls into.
condensed (optional) DEPRECATED Use condense instead. Support for this will be removed entirely in a future release.
transform (optional) Optionally process each row of data. Supports [mean, count, min, max, sum], or any callable that accepts a list of datapoints according to the type of series (e.g histograms are dictionaries, counts are integers, etc). Transforms are called after read_func has cast the data type and after resolution data is optionally condensed. If transform is one of (list,tuple,set), will load the data once and run all the transforms on that data set. If transform is a dict of the form { transform_name : transform_func }, will run all of the transform functions on the data set.
fetch (optional) Function to use instead of the built-in implementations for fetching data. See Customized Reads.
process_row (optional) Can be a callable to implement Customized Reads.
condense (optional) If using resolutions, True will collapse the resolution data into a single row. Can be a callable to implement Customized Reads.
join_rows (optional) Can be a callable to implement Customized Reads.
collapse (optional) True will collapse all of the data in the date range into a single result. Can be a callable to implement Customized Reads.
Returns an ordered dictionary of { interval_timestamp : { resolution_timestamp: data } }, where interval_timestamp and resolution_timestamp are Unix timestamps and data is a data structure corresponding to the type of series, or whatever transform returns. If not using resolutions or condensed=True, the dictionary will be of the form { interval_timestamp : data }.
All variations of transform and the resulting format of data are the same as in get.
If both start and end are defined, the returned data will start and end on intervals including those timestamps. If only start is defined, then the return data will start with an interval that includes that timestamp, with the total number of intervals returned defined by steps. If only end is defined, then the return data will end with an interval that includes that timestamp, with the total number of intervals preceeding it defined by steps.
It is important to note that the interval timestamps in the returned data will not necessarily match start or end. This is because of the consistent hashing scheme that kairos uses, such that start and end will be translated into the bucket in which it can be found.
iterate
Almost identical to get except it does not accept a timestamp argument.
name The name of the statistic, or a list of names whose data will be joined together.
interval The named interval to read from
transform (optional) Optionally process each row of data. Supports [mean, count, min, max, sum], or any callable that accepts datapoints according to the type of series (e.g histograms are dictionaries, counts are integers, etc). Transforms are called after read_func has cast the data type and after resolution data is optionally condensed. If transform is one of (list,tuple,set), will load the data once and run all the transforms on that data set. If transform is a dict of the form { transform_name : transform_func }, will run all of the transform functions on the data set.
fetch (optional) Function to use instead of the built-in implementations for fetching data. See Customized Reads.
process_row (optional) Can be a callable to implement Customized Reads.
condense (optional) If using resolutions, True will collapse the resolution data into a single row. Can be a callable to implement Customized Reads.
join_rows (optional) Can be a callable to implement Customized Reads.
Returns a generator which iterates over ( timestamp : data ) tuples, where timestamp is a Unix timestamp and data corresponds to the rules documented in get.
Customized Reads
ALPHA This feature is still being explored and the API may change significantly.
There are times when the data in a timeseries requires processing to be pushed onto the datastore.
There are times when one needs custom control over the reading and processing of data in a timeseries. As there is no good way to do this generically, the get and series API supports several keyword arguments to customize access to the data. Common use cases are to handle large sets of data that can be processed in the datastore, and situations where one wants to implement cutom analysis of the dataset such as calculating variance.
General
The following functions can be overloaded with keyword parameters to get and series (collapse being only used for a series).
fetch
A customized database read function. The usage varies depending on the backends which are described in detail below. IMPORTANT You are welcome to change the type of the return value, but be wary that transforms, condense and collapse functionality may not work properly with the changed data types.
process_row
The function which handles the type casting of the data read from the backend and also calling the read_func if it has been defined for the time series. It is required that you define this function if you overload fetch such that the returned data type is not the same as the time series’ native format (dict for histogram, list for series, etc).
The function must be in the form of process_row(data), where:
data The row data generated by the native or fetch implementation, not including any time stamps.
The function may return any data type, but if it’s not the native format of the time series, additional downstream functions may have to be overloaded.
condense
If the condense argument is a callable, the caller can override how resolution data is collapsed (reduced) into a single interval. The argument will always be in the form of:
{
'resolution_t0' : <data_t0>,
'resolution_t1' : <data_t1>,
...
'resolution_tN' : <data_tN>,
}
Where <data_tN> is the data returned from the native or fetch implementation and passed through the native or custom process_row implementation.
The function should return a single value, optionally in the same format as <data_tN>, but this method could also be used for calculating such things as rate of change or variance within a time interval.
join_rows
If the join_rows argument is a callable and the name parameter to get or series is one of (list,tuple,set), this method will be called to join the data from several named timeseries into a single result. The argument will always be in the form of:
[ <data_series0>, <data_series1>, ... <data_seriesN> ]
Where <data_series0> will be the data within a single timestamp window in the series’ native format or whatever was generated by custom implementations of fetch, process_row and/or condense. It is important to note that not every series will contain data points within a given time interval.
In addition to reducing multiple time series’ worth of data within an interval into a single result, this method could be used to implement cross-series analytics such as unions, intersections and differentials.
collapse
If the collapse argument is a callable, the caller can override how interval data is collapsed (reduced) into a single result. The native implementation is to call the condense function implemented by a time series. The arguments are the same as a custom condense function, as-is the expected return value.
It’s important to note that if collapse is defined, the series will automatically be condensed as well, so if fetch is overloaded to return a custom data type, then condense must also be defined. If collapse is True, the custom condense function will be used if defined.
In addition to collapsing the result of a time series into a single data set, this method could also be used to calculate data across a time series, such as variance.
transform
As noted previously, transform can be any callable, list of names or callables, or a named map of transform names or callables. The transforms will be processed after all previous native or custom read functions, including collapse.
Redis
The function must be in the form of fetch(handle, key), where:
handle Either a Redis client or pipeline instance
key The key for the timeseries data
The return value should correspond to the data type of timeseries, e.g. dict for a histogram. One should always assume that handle is both a pipeline and a client, and fetch should return the result of, e.g. handle.hlen(...), but that it cannot be used to return a literal, such as lambda: h,k: { 'foo' : h.hlen(k) }
Mongo
The function must be in the form of fetch(handle, **kwargs), where:
handle A PyMongo Collection
spec The (suggested) query specification
sort The (suggested) sort definition for the query
method The suggested method to use on the handle
The required return value depends on the value of method.
find_one Should return a hash in the form { value : <data> }, where <data> should correspond to the data type of the timeseries, e.g. list for a series. May directly return a result from pymongo.collection.find_one.
find Should return an iterable in the form [ { value: <data> }, ... ], where <data> follows the same rules as find_one.
Re-implementing the default functionality would look like:
def mongo_fetch(handle, spec, sort, method):
if method=='find':
return handle.find( spec=spec, sort=sort )
elif method=='find_one':
return handle.find_one( spec )
SQL
The function must be in the form fetch(connection, table, name, interval, i_start, i_end), where:
connection A SQLAlchemy Connection
table A SQLAlchemy Table
name The name of the stat to fetch
interval The interval of the stat to fetch
i_start The interval timestamp (starting) key
i_end (optional) For a series, the ending timestamp key
The return value should be in the form of
{
'interval_t0' : {
'resolution_t0t0' : <data_t0t0>,
'resolution_t0t1' : <data_t0t1>,
...
'resolution_t0tN' : <data_t0tN>
},
'interval_t1' : { ... },
...
'interval_tN' : { ... },
}
If the series doesn’t use resolutions, then resolution_tNtN should be None, and so each interval will be in the form { 'interval_tN: { None : <data_tN> } }. This is inherent in the way that data is stored within the tables.
If i_end is supplied, the query should be over the range i_time >= i_start AND i_time <= i_end, else the query should be for the interval i_time = i_start.
Cassandra
The function must be in the form fetch(connection, table, name, interval, i_start, i_end), where:
cursor A cql Connection
table The name of the table
name The name of the stat to fetch
interval The interval of the stat to fetch
i_start The interval timestamp (starting) key
i_end (optional) For a series, the ending timestamp key
The return value should be in the form of
{
'interval_t0' : {
'resolution_t0t0' : <data_t0t0>,
'resolution_t0t1' : <data_t0t1>,
...
'resolution_t0tN' : <data_t0tN>
},
'interval_t1' : { ... },
...
'interval_tN' : { ... },
}
If the series doesn’t use resolutions, then resolution_tNtN should be None, and so each interval will be in the form { 'interval_tN: { None : <data_tN> } } and can be determined when a row has an r_time of -1.
If i_end is supplied, the query should be over the range i_time >= i_start AND i_time <= i_end, else the query should be for the interval i_time = i_start.
Deleting Data
There are two methods to delete data.
delete
Takes a single argument, the name of the timeseries. Will delete all data for that timeseries in all intervals.
delete_all
Deletes every timeseries for all intervals. This method may be fast in data stores that support optimized deletes, else it will have to delete for each timeseries returned in list.
expire
Takes a single argument, the name of the timeseries. For storage engines that do not support expiry, such as SQL, will delete expired data from intervals for which steps is defined. All other storage engines will raise the NotImplementedError exception.
Dragons!
Kairos achieves its efficiency by using Redis or Mongo TTLs and data structures in combination with a key naming scheme that generates consistent keys based on any timestamp relative to epoch. However, just like RRDtool, changing any attribute of the timeseries means that new data will be stored differently than old data. For this reason it’s best to completely delete all data in an old time series before creating or querying using a new configuration.
Installation
Kairos is available on pypi and can be installed using pip
pip install kairos
If installing from source:
with development requirements (e.g. testing frameworks)
pip install -r development.pip
without development requirements
pip install -r requirements.pip
Note that kairos does not by default require the redis package, nor does it require hiredis though it is strongly recommended.
Tests
Use nose to run the test suite.
$ nosetests
The test suite can be controlled through several environment variables, all defaulting to true.
TEST_REDIS true
TEST_MONGO true
TEST_SQL true
TEST_CASSANDRA true
TEST_SERIES true
TEST_HISTOGRAM true
TEST_COUNT true
TEST_GAUGE true
TEST_SET true
SQL_HOST sqlite:///:memory:
CASSANDRA_KEYSPACE kairos
Roadmap
Batch inserts
Round-robbin intervals for datastores without TTLs
Round-robbin databases: memcache (and compatible, e.g. ElastiCache), Riak, DynamoDB, SimpleDB, GDBM, Berkeley DB, and more
Redis optimizations
Capped collection support for mongo
InfluxDB support
Bloom filters
Joined series populate a data structure at query time
Joined series support concurrency “runner”
License
This software is licensed under the New BSD License. See the LICENSE.txt file in the top distribution directory for the full license text.
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
